Multi-turn GPT Image 2.5 editing runs only through OpenAI's Responses API

On the Image API you call client.images.generate with model set to gpt-image-2.5-sunburst, and the image comes back as base64 in result.data[0].b64_json ^[6]. On the Responses API you set a mainline model at the top level, the docs example uses gpt-6-astra, put the image model in the image_generation tool's model field, and then filter response.output for items whose type is image_generation_call ^[7][5]. Two model strings per request instead of one ^[22].

Single-shot editing works on both. The Image API has an Edits endpoint that modifies an existing image partially or entirely ^[2]. The Responses API adds iteration: multi-turn editing, and image File IDs accepted as inputs rather than only bytes ^[3]. Turn state comes from passing the image generation call output back in context, or from previous_response_id ^[8]. The guide's routing advice is direct: one image from one prompt, use the Image API; conversational, editable experiences, use the Responses API ^[4].

The tool's action parameter decides behaviour at runtime. Setting it to auto lets the model choose whether to generate a new image or edit one already in the conversation, generate always creates, and edit forces an edit when an image is in context ^[9]. Force edit with nothing in context and the call returns an error ^[10]. The guide documents that error as the expected result.

The docs do not settle pricing. The New Stack reports identical token rates for the two models ^[14], and OpenAI's note says "Token rates match GPT Image 2. The GPT Image 2 calculator does not estimate GPT Image 2.5 token consumption." ^[15]. According to The New Stack, there is no published guidance for estimating 2.5 consumption at all ^[16]. Image output at $30 per million tokens is 3 cents per thousand output tokens, so a 1,000-token render costs 3 cents and a 4,000-token one costs 12 cents ^[25]. Which of those you get comes out of a metered test run ^[21].

Treat the latency figure the same way. OpenAI says Flare delivers higher-quality images than GPT-Image-2 with 50% lower latency ^[17], and calls it "the default choice for most applications" ^[18]. That 50% bounds your own improvement only if GPT-Image-2 is what you are replacing, at your sizes, quality settings and prompt mix. Sunburst, which OpenAI describes as "built for premium visual workflows that benefit from tighter control across edits", comes with "longer generation times" per the launch announcement ^[19]. Per The New Stack, neither the announcement nor the model pages quantify that gap ^[20].

Quality, size, format and compression are adjustable from both APIs ^[12], so output settings do not decide this. The question is whether the product will ever want a second edit of the same image. If it might, write against the Responses API shape first, because switching later changes where the model is named and how the response is read ^[5][7].